The margins are moving the wrong way
I’ve been watching Anthropic’s numbers for a while now, and something doesn’t add up. TechCrunch reported this week that at a moment when Anthropic’s compute margins should be getting better, they’re getting worse. That’s not a rounding error. That’s a signal.
Let me be blunt: we’ve spent two years listening to every AI lab on earth promise that scale would solve the economics. Bigger clusters, cheaper inference, better chips, and suddenly the unit economics would flip. That was the story. The pitch deck version of gravity. And yet here we are in late 2026, and one of the two most important frontier labs on the planet is watching its compute margins deteriorate exactly when they were supposed to improve.
What struck me here isn’t the headline. It’s the timing. If this were 2024, I’d shrug. Everyone was burning cash to buy market position. But 2026 is supposed to be the year the industry grows up. The year the spreadsheet stops being a punchline.
Why compute costs don’t behave like normal costs
Here’s the thing people outside the industry keep getting wrong. Compute isn’t like cloud storage. You don’t just wait for prices to fall and watch your margins expand. Frontier model training and inference are locked in a kind of arms race where the cost per token falls, but the number of tokens you need to stay competitive rises faster. You’re on a treadmill that speeds up every quarter.
Anthropic’s problem, as I read it, is structural. The company has staked its identity on safety and on being the reliable, enterprise-grade alternative to OpenAI. That’s a smart positioning. Enterprises pay for reliability. But reliability at the frontier is expensive. You can’t run a lean inference stack when your customers expect Claude to hold up under load, handle long context, and not hallucinate its way through a legal contract.
So you buy more compute. You buy it at spot prices when you can, you sign long-term contracts when you must, and you pray the depreciation curve is kind. But the depreciation curve is never kind at this scale. GPUs age fast. The H100s that felt like magic two years ago are now the thing you’re trying to sweat before the next generation makes them look quaint.
The enterprise trap
I think there’s a trap here that nobody wants to name out loud. Enterprise AI is a margin business that looks like a software business but behaves like a utility. Your customers want SLAs. They want uptime. They want predictable latency. All of that costs money, and none of it scales the way SaaS margins scaled in the 2010s.
Compare that to the consumer side. ChatGPT can afford to be sloppy at the edges because a consumer who gets a weird answer just… asks again. An enterprise customer with a compliance team does not ask again. They file a ticket. They escalate. They renegotiate the contract.
Anthropic has chosen the harder road on purpose, and I respect the choice. But respect doesn’t pay for compute. The company is now in the position of needing to raise prices, tighten usage limits, or find a compute partner willing to eat some of the margin. None of those are fun conversations.
What the hyperscalers are really selling
This is where the story gets interesting. The reason Anthropic is grappling with compute economics is that the hyperscalers have quietly become the most important players in the AI value chain, and they know it. AWS, Google, and Microsoft aren’t just infrastructure providers anymore. They’re the landlords of the entire frontier.
And landlords always get paid first.
I’ve argued for years that the AI labs are renting their future from companies that would happily build a competing product if the rent stopped coming in. That’s not paranoia. That’s just reading the org chart. Google funds Anthropic and also runs Gemini. Amazon funds Anthropic and also has its own models. The checks clear, but the strategic ambiguity is the point.
So when Anthropic’s compute margins worsen, the question isn’t just « can they fix it? » The question is « who benefits from them not fixing it? » I’ll let you draw your own lines there.
The inference bill nobody wants to talk about
Training gets the headlines. Inference pays the bills, or fails to. Every time you send a long document to Claude and ask it to summarize, you’re triggering a chain of computation that costs real money. Multiply that by millions of enterprise seats and the numbers get ugly fast.
There’s a reason every lab is obsessed with smaller, cheaper models right now. It’s not because small models are cooler. It’s because the big ones are financial liabilities at scale. Anthropic knows this. That’s why you’re seeing more emphasis on model routing, caching, and distillation across the industry. The goal is to serve the same quality of answer for a fraction of the tokens.
But here’s the catch: every efficiency gain gets competed away. If Anthropic figures out how to cut inference costs by 40%, OpenAI does too, and then the market expects 40% cheaper pricing, and the margin improvement evaporates. That’s the treadmill again. You run faster just to stay in place.
Is this a crisis or just a phase?
I want to be careful here because the temptation is to write a doom piece. « Anthropic is bleeding money, the AI bubble is bursting, everyone panic. » That’s lazy. The more interesting read is that this is what the maturation of an industry actually looks like.
Every infrastructure-heavy sector goes through this. Railroads, telecom, cloud computing. The early years are about building capacity and the middle years are about figuring out who can actually operate profitably. The companies that survive aren’t the ones with the biggest clusters. They’re the ones with the best cost discipline and the clearest customer lock-in.
Anthropic has the customer lock-in, at least on the enterprise side. What it doesn’t yet have is the cost discipline. The company’s culture is research-first, safety-first, and that’s admirable. But research-first cultures tend to treat the finance team as a necessary annoyance rather than a strategic partner. That has to change if the margins are going to change.
What I’d watch next
Three things. First, pricing changes. If Anthropic quietly raises enterprise rates or restructures its usage tiers in the next two quarters, that tells you the margin pressure is real and they’re passing it on. Second, compute partnerships. Any new deal that looks like a revenue share or a co-investment rather than a straight purchase is a sign they’re trying to offload risk. Third, model mix. If Claude’s flagship gets leaner and the company pushes customers toward cheaper tiers, that’s the efficiency play in action.
What I wouldn’t watch is the press release language. Every lab will tell you the economics are fine and the future is bright. The spreadsheet tells a different story, and the spreadsheet is the one that eventually gets a vote.
The uncomfortable truth about AI compute economics
Here’s my honest take. The AI industry has spent years pretending that compute is a scaling problem. It’s not. It’s a business model problem. You can throw as many GPUs at the problem as you want, but if the revenue per token doesn’t exceed the cost per token by a healthy margin, you’re just running a very expensive science experiment with a nice demo layer on top.
Anthropic is one of the smartest teams in the business. If anyone can thread this needle, they can. But the TechCrunch report should be a wake-up call for the whole sector. The era of « growth at any cost » is ending, and the companies that survive will be the ones that figured out how to make compute boring. Profitable, predictable, boring compute.
That’s not the story anyone wanted to tell in 2023. It’s the only story that matters in 2026.
Original source: read the full article